[Pilot invitation] WFA2 affine and ends-free differential validation

[logannye]

@smarco — we built hcp-align, a pure-Rust exact alignment oracle that independently verifies objective values, spans, and canonical =XID CIGARs. The development tooling includes a pinned WFA2-lib v2.3.6 adapter with heuristics disabled.

Would you be willing to spend 30–60 minutes trying one small affine or ends-free corpus? The useful review is narrow: catch score-translation mistakes, invalid converted paths, or cases where different optimal CIGARs are incorrectly treated as failures. Every converted result is independently rescored before comparison.

A pilot counts as either 100 real records, a complete smaller failure corpus, or one reproducible backend discrepancy. We will perform reduction and add any confirmed defect as a regression fixture.

Pilot guide: https://github.com/logannye/hcp-dp/blob/main/docs/pilot-guide.md
Signed prerelease: https://github.com/logannye/hcp-dp/releases/tag/v0.2.0-alpha.2

Sequence data may stay private; minimized fixtures, digests, aggregate statuses, and scoring metadata are sufficient. This is a request for technical feedback, not an endorsement. If the timing or scope is not useful, no response is expected.

[logannye]

@smarco — one brief follow-up in case the original pilot ask was too broad. The smallest useful contribution would be a quick confirmation or correction of our WFA2 affine/ends-free score translation; no corpus or endorsement is required. We can provide a ready-to-run fixture and handle all independent verification and reduction. If this is not a fit, no reply is needed and we will not follow up again.

[smarco]

Hi @logannye,

If that helps, we will do our best to support you. Let us know what you need from our end, and we'll do our best to help.
Feel free to reach out via email too.

Best,

[logannye]

Thank you, @smarco — that is very helpful. The smallest concrete ask is to sanity-check the WFA2 v2.3.6 semantics behind this translation; a short correction or confirmation is enough to guide us.

For HCP's maximize-score model, where a match scores m, a mismatch scores -x, a linear gap base scores g < 0, and an affine gap of length L scores go + L*ge, we send WFA:

• mismatch: 2 * (m + x)
• linear indel: m + 2 * abs(g)
• affine opening: 2 * abs(go)
• affine extension: m + 2 * abs(ge)

We recover the alignment score as (m * (query_length + aligned_target_length) - WFA_cost) / 2, rejecting non-positive translated costs and non-integral recovered scores. We also normalize WFA D (pattern/query consumption) to SAM I, and WFA I (text/target consumption) to SAM D. For semiglobal alignment, we enumerate target substrings and use WFA end-to-end for each candidate because the score-to-cost baseline depends on aligned target length.

Is any part of that inconsistent with WFA2's affine or CIGAR semantics?

If running the pinned contract is easier, this is the exact audited setup:

git clone https://github.com/logannye/hcp-dp.git
cd hcp-dp
git checkout 61a945e2fdd00020cb7ca3b685cf87860c6037ab
git clone --filter=blob:none https://github.com/smarco/WFA2-lib.git ../WFA2-lib
git -C ../WFA2-lib checkout bcf473a6561fa297934a80eaa7c04b4ee412360c
scripts/build-wfa2-adapter.sh ../WFA2-lib "$PWD/target/wfa2-adapter"
cargo run --locked --manifest-path tools/differential/Cargo.toml -- contract \
  --left internal:auto --right "wfa2:$PWD/target/wfa2-adapter" \
  --max-length 3 --output differential-report.json
python3 scripts/check_differential_allowlist.py --report differential-report.json

Our hosted run produced 900 independently verified, objective-equivalent cases; alternate optimal CIGAR and operation-count differences were retained as diagnostics rather than failures.

If you have one affine/ends-free boundary pair you consider especially diagnostic, sequence text can remain private: scoring parameters plus a minimized synthetic equivalent or digest/status summary is sufficient. We will run exact verification, reduce any disagreement, and post the result here.